Adaptive feature consolidation residual network for exemplar-free continuous diagnosis of rotating machinery with fault-type increments

Models employing the continual learning (CL) paradigm hold promising potential for application in rotating machinery fault diagnosis. These models allow the diagnostic network to continuously learn and update in dynamically changing data scenarios, eliminating the need for laborious retraining processes. However, most existing replay-based approaches require the model to re-access the previous task data, thereby increasing additional computation and memory occupation. In addition, CL-based models exhibit high plasticity, resulting in feature drift and posing challenges for effective model compensation. To overcome these limitations, a novel adaptive feature consolidation residual network (AFCRN) is proposed for exemplar-free continuous diagnosis of rotating machinery with fault-type increments. Specifically, pseudo-metrics induced by the empirical feature matrix in the feature space consolidate the feature representation by regularizing the drift, which is highly correlated with the previous diagnostic task. Gaussian prototypes are employed to mitigate bias induced by similar diagnostic tasks. Moreover, a new asymmetric cross-entropy loss function is utilized to enable the model classifier to adapt to a continuously updated backbone network, thereby balancing new data with prototypes. Finally, the effectiveness of the proposed method is evaluated through two case studies of continuous fault diagnosis. The experimental analysis demonstrates that the designed AFCRN offers significant advantages in fault diagnosis of fault-type increment scenarios.